Meta Chief Scientist LeCun: The new generation of AI architecture will be released within 3 to 5 years, with robotics technology as the core

Meta's Chief AI Scientist Yann LeCun stated on Thursday that a "completely new AI architectural paradigm" will emerge within the next 3 to 5 years, with capabilities far surpassing existing AI systems. He also predicted that the coming years could become "the decade of robotics," where advancements in AI and robotics will combine to unlock a new generation of intelligent applications.

Media reports indicate that during a session titled "Technology Debate" at the Davos Forum, LeCun pointed out that the current "form of AI," namely generative AI and large language models (LLMs), is not as powerful as it seems. He believes that while these models are useful, they still have shortcomings in many areas.

“I think the current LLM models have a relatively short lifespan, possibly only three to five years. I believe that within five years, no one will consider the current models as core components of AI systems. I think we will see the emergence of a completely new AI architectural paradigm that may not be constrained by current AI systems.”

He stated that the "limitations" of current AI models hinder machines from achieving true intelligent behavior, which can be attributed to four key reasons: a lack of understanding of the physical world, a lack of persistent memory, a lack of reasoning ability, and a lack of complex planning capabilities.

“LLMs are actually unable to achieve these. Therefore, in the coming years, we will witness another revolution in AI. We may need to rename it because it may no longer be the generative AI we understand today.”

World Models

This viewpoint echoes what LeCun has expressed in the past. Several AI leaders, including LeCun, have proposed the so-called "world model," which aims to help machines understand the dynamics of the real world, including memory, common sense, intuition, and reasoning abilities, far beyond the capabilities of current systems, which mainly excel at pattern recognition.

Previously, LeCun indicated that this might take about 10 years, but the current estimate has brought the timeline closer. However, it remains unclear what level can be achieved within this timeframe.

“LLMs are good at processing language but not at thinking. So we are working on developing systems that can build world mental models. If our current plans succeed according to the expected timeline, then in three to five years, we will have a system of a completely different paradigm. They may possess some degree of common sense. They may be able to learn how the world works by observing the world and even interacting with it.”

The Decade of Robotics

Although generative AI is impressive, capable of passing the bar exam or discovering new drugs, LeCun believes that robotics may become the core of the next wave of AI applications, especially in practical scenarios.

Currently, Meta is conducting some research in the field of robotics, and OpenAI, the developer of the currently popular AI star ChatGPT, is also exploring robotics. Earlier this month, it was reported that OpenAI has formed a new robotics team focused on "general," "adaptive," and "multi-functional" robots, aiming to achieve human-like intelligence and adapt to the real world LeCun stated:

"We currently do not have robots that can understand the physical world like cats do—cats' understanding of the physical world far exceeds anything we can achieve with AI. Perhaps the next decade will be the decade of robots, and we may have sufficiently intelligent AI systems that can understand how the real world operates."